The web was designed to remember. That is its greatest achievement and its deepest pathology.
We built a global system for sharing knowledge, and it works — beautifully, catastrophically well. A page published in 1997 is still accessible in 2026. A draft email sent by accident lives in someone's inbox forever. A post written in anguish at 3am becomes the first result in a search for your name a decade later.
We have created an infrastructure of permanent text.
And on top of it, we have built a culture that treats all content as if it were meant to be permanent — every draft a publication, every sketch a final canvas, every midnight thought a considered position.
The web's protocol stack was designed for retrieval, not for nuance. HTTP provides rich semantics for content type, caching, and authorization. But it provides no vocabulary for the ontological status of content. No mechanism by which a creator may assert that a work is, by intention, without title, without final form, without claim to permanence.
The question "what is the epistemic status of this content?" is not one HTTP was designed to answer.
Every UX pattern on every publishing platform is an optimization for completion: the progress bar, the word count, the publish button. We have built infrastructure architecturally hostile to the draft, the sketch, the experiment, the thought at midnight.
They are not abstract. They appear every day.
Search engines index draft documents as authoritative positions. A writer publishes a working hypothesis, notes it as "thinking out loud," and finds it cited in papers two years later as a concluded argument. The draft becomes the record.
AI training pipelines ingest ephemeral, context-dependent content as if it were stable knowledge. A conversation at 2am, a grief post, a private forum thread scraped without context — these become the foundations of systems that speak with the authority of the web's collective record. The temporary becomes the permanent.
Archival services preserve content whose creators intended it to vanish. Not maliciously — they are doing exactly what they were built to do. A web built for permanence does what a web built for permanence does.
Content carries context. When you write something, you write it for a particular audience, in a particular moment, with particular expectations about how it will be read. A note to a collaborator is not a press release. A working draft is not a published paper. A post to a small forum is not a deposition.
The web erases that context. A URL is a URL. It presents identically to your colleague, to a stranger in another country, to a journalist writing a story ten years later, and to a language model ingesting the entire indexed web at once. There is no encoding for "this was written for twenty people who already understand the situation."
This is what researchers call context collapse: the flattening of distinct, intended audiences into a single undifferentiated public. It is not a bug anyone introduced; it is a structural consequence of a system built for retrieval, not for nuance. The web was never designed to carry the author's intent — only the author's text.
You cannot give a URL an audience. You can, however, give it an epistemic status. You can say: this is unfinished, this is temporary, this belongs to no one in particular. That assertion can travel with the URL, in a form that machines can parse and act on. That is what the Untitled Protocol adds to the web's vocabulary.
In 1994, Martijn Koster wrote a text file called robots.txt.
Seventeen lines. It said: here is where you may and may not crawl. Voluntary.
No enforcement mechanism. No legal authority.
The search industry adopted it because compliance was cheaper than the alternative —
lawsuits, hostile legislation, reputational damage. Today, every major crawler
checks robots.txt before indexing. A text file, adopted by social
consensus, became a foundational layer of how the web works.
But robots.txt is binary: crawl or don't crawl. It cannot express
"index this, but note it as a draft." It cannot express "this content is public,
but its author is anonymous by choice." It cannot express "this page is live now,
but was not meant to be permanent." The vocabulary that covers
what kind of thing this is does not exist in robots.txt.
That vocabulary has never existed anywhere in a machine-readable form. Until now.
The Untitled Protocol is the next text file.
It does not fight the web's memory. Fighting the web's memory would be like fighting the tide. Instead, it extends the web's vocabulary. It adds, to the existing grammar of HTTP and HTML, a new kind of sentence — one that says this thing exists, but it is not done, and it does not want to be remembered as if it were.
The generative AI industry is attempting to navigate an ethical catastrophe — the ingestion of human creativity without consent — with legal arguments instead of technical solutions. This is a mistake, and it is a fixable one.
robots.txt taught us that the right answer to "can we crawl this?"
was not a legal framework but a machine-readable standard. The Untitled Protocol
applies the same architecture to the training data consent problem. It does not
restrict AI. It gives AI systems the information they need to make ethical choices
at scale.
A company that implements it can, for the first time, make a technically verifiable claim: "We do not train on content that its creator declared as ephemeral or anonymous."
The problem is not that AI companies are evil. The problem is that they don't have a signal.
We built the signal.
The condition of being without a final title is not a failure state.
It is a legitimate, affirmative declaration about the relationship between content and the world. A work can be present without being permanent. It can exist without being attributed. It can be shared without being archived.
The Untitled Protocol grants creators the ability to declare, in language that machines can read and respect, that a given piece of content belongs to the present moment, to no one in particular, and to a state of becoming rather than being.
That sentence has never existed in machine-readable form before. You could write it in English, in an author's note, in a disclaimer. But search engines don't read disclaimers. AI training pipelines don't parse authorial intent. The web is increasingly read by machines, and machines only understand what you say in their language.
The Untitled Protocol says, in the language that machines read:
X-Untitled: nowNot then. Not always. Now.
And now is enough.